Very Large Cross-lingual Resources at OAEI 2008 Laura Hollink Véronique Malaisé Vrije Universiteit Amsterdam

Task description Mappings between three resources: – GTAA Thesaurus of the Dutch Institute for Sound and Vision. 4 Facets: Subject, Location, People, Name – WordNet Lexical database – DBPedia ‘structured version of Wikipedia’ 3 sets of mappings exactMatch, broadMatch, narrowMatch.

Rationale Different languages – Archive with GTAA metadata in Dutch only – Broaden user group – Integrated access to archives of other countries? Large resources – Disambiguation becomes serious problem Heterogeneous resources – Different structure – Weak or inconsistent structure – Large parts have no counterparts, when to stop mapping?

Cross-Lingual GTAA in Dutch: – Preferred labels – Alternative labels – Scope notes WordNet in English – Word-senses – Glosses DBPedia in English and (most of the time) Dutch – Titles – Abstracts

Different Schema’s GTAA in SKOS – Skos:concepts with pref- and altLabels – Narrower/Broader relations between the concepts WordNet – Synsets with word-senses – Hyponym relations between synsets. DBPedia – Things with titles and abstracts – links to dbpedia categories, rdf:type links to yago classes – Hierarchical structure of yago classes and categories.

Results One Participant: DSSIM Vocabulary#Concepts#Mappings to In VocabularyWordNetDBPediaGTAA WordNet82,000NA28,9742,405 DBPedia2,700,00028,974NA13,156 GTAA160,0002,40513,156NA Subject3, ,363NA People97,000822,238NA Name27, ,989NA Location14, ,566NA

Evaluation Process Precision – GTAA-WordNet & GTAA-DBPedia Inspection of 100 mappings per GTAA facet. – WordNet-DBPedia Inspection of 100 mappings – Correct or Incorrect or Narrow/Broader/Related Recall – GTAA-WordNet & GTAA-DBPedia Comparison to a reference alignment of 100 GTAA People and 100 GTAA Subjects

Pre-existing WN-DBPedia Mappings Type Links – Air_New_Zealand wordnet-type synset-airline-noun-2 – No overlap between DSSIM mapping and wordnet-type links. – DSSIM mappings of things with a wordnet-link performed less than DSSIM mappings of things without a wordnet-link. Yago – most DBPedia "things" are instances in the YAGO ontology. – Dbpedia categories are classes in the YAGO ontology, subclasses of wordnet synsets. – “Crazy”_Joe_Davola rdf:type FictionalCharacter Overlap between DSSIM results and Yago? We are looking (mainly) for exactMatches, not type links.

Conclusions Also other types of links than exactMatch are necessary: – Subject:pausbezoeken -> List_of_pastoral_visits_of_Pope_John_Paul_II_out side_Italy. – Location:Venezuela -> synset-Venezuelan-noun-1 – Subject:Verdedigingswerken -> fortification

Conclusions Context would help a lot. GTAA facet information – Person:GoghVincentvan -> synset-vacationing-noun-1 – Location:Harlem -> synset-hammer-noun-8 – Location:Melbourne -> synset-Melbourne-noun-1 Titles of DBPedia ‘referring pages’ used as alternative labels. Person:SummerGordon -> Sting_(musician) But: no longer a generally applicable tool.

Reflection on the evaluation process Disambiguation of DBPedia/WordNet concepts very hard, also for evaluator. – Subject:leguanen -> Iguana or Iguanidae? – Multiple mappings are reasonable. When is a mapping ‘Related’? DBPedia disambiguation pages. GTAA contains may Dutch-specific concepts – Diogenes = TV program Take into account the confidence measures

Input from you What do OAEI participants think of this task? How can we improve it?

Why Not? Too much preprocessing required? The three resources have different schema’s? Tool was not built for resources that large? Tool was not built for multi-lingual matching?

Thank you!